Apparatus and method for detecting images within spam

- Proofpoint, Inc.

A method is described that includes comparing a characteristic of an image to stored characteristics of spam images. The method also includes generating a signature of the present image. The method further includes comparing the signature of the present image to stored signatures of spam images. The method also includes determining the spam features corresponding to the stored signatures of spam images that match the signature of the present image.

Skip to: Description  ·  Claims  ·  References Cited  · Patent History  ·  Patent History
Description
CLAIM TO PRIORITY

This application is a divisional of U.S. patent application Ser. No. 11/652,715, filed on Jan. 11, 2007, now U.S. Pat. No. 8,290,203, all of which are hereby incorporated by reference in their entirety into this application.

BACKGROUND OF THE INVENTION

Field of the Invention

This invention relates to electronic message analysis and filtering. More particularly, the invention relates to a system and method for improving a spam filtering feature set.

Description of the Related Art

“Spam” is commonly defined as unsolicited bulk e-mail, i.e., email that was not requested (unsolicited) and sent to multiple recipients (bulk). Although spam has been in existence for quite some time, the amount of spam transmitted over the Internet and corporate local area networks (LANs) has increased significantly in recent years. In addition, the techniques used by “spammers” (those who generate spam) have become more advanced in order to circumvent existing spam filtering products.

Spam represents more than a nuisance to corporate America. Significant costs are associated with spam including, for example, lost productivity and the additional hardware, software, and personnel required to combat the problem. In addition, many users are bothered by spam because it interferes with the amount of time they spend reading legitimate e-mail. Moreover, because spammers send spam indiscriminately, pornographic messages may show up in e-mail inboxes of workplaces and children—the latter being a crime in some jurisdictions.

Spam filters attempt to remove spam without removing valid e-mail messages from incoming traffic. For example, spam filters scan email message headers, metatag data, and/or the body of messages for words that are predominantly be used in spam, such as “Viagra” or “Enlargement.” Current email filters may also search for images which are known to be used in spam messages. Hashing algorithms such as MD5 are used to generate image “fingerprints” which uniquely identify known spam images.

Over the years, spammers have become more creative in disguising their messages, e-mails, or advertisements as legitimate incoming traffic to avoid detection by spam filters. Specifically, spammers typically obfuscate words which would normally be identified by spam filters. For example, “Viagra” may be spelled “V!agra” or “Enlargement” may be spelled “En!@rgement.” With respect to images, spammers often embed random data within spam images to modify the image fingerprint, and thereby avoid detection.

Thus, improved mechanisms for detecting obfuscated images within email messages are needed.

BRIEF DESCRIPTION OF THE DRAWINGS

A better understanding of the present invention can be obtained from the following detailed description in conjunction with the following drawings, in which:

FIG. 1 illustrates a message filtering apparatus.

FIG. 2 illustrates an example method of the image analyzer of the message filtering apparatus.

FIG. 3 illustrates the method of the Fuzzy Feature Detection Algorithm (Fuzzy80) described in FIG. 2 and executed by the image analyzer of the message filtering apparatus of FIG. 1.

FIG. 4 illustrates one example image cropping method.

FIG. 5 illustrates another example image cropping method.

FIG. 6 illustrates the lookup table for the Fuzzy80 algorithm of FIG. 3 to determine features of a present image.

FIG. 6A illustrates a specific example of comparing two vectors during the Fuzzy80 algorithm of FIG. 3.

FIG. 7 illustrates the method of the GIF portion/algorithm (GIF80) of the GIF/JPEG Feature Detection Algorithm (GIF80/JPEG80) described in FIG. 2 and executed by the image analyzer of the message filtering apparatus of FIG. 1.

FIG. 8 illustrates the typical format of a GIF.

FIG. 9 illustrates the lookup table for the GIF80 algorithm of FIG. 7 to determine features of the present image.

FIG. 10 illustrates the method of the JPEG portion/algorithm (JPEG80) of the GIF/JPEG Feature Detection Algorithm (GIF80/JPEG80) described in FIG. 2 and executed by the image analyzer of the message filtering apparatus of FIG. 1.

FIG. 11 illustrates the lookup table for the JPEG80 algorithm of FIG. 10 to determine features of the present image.

FIG. 12 illustrates a combined lookup table for the GIF80 algorithm and the JPEG80 algorithm, combining the lookup tables illustrated in FIGS. 9 and 11.

 DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

Described below is a system and method for detecting images used in spam. Throughout the description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present invention. It will be apparent, however, to one skilled in the art that the present invention may be practiced without some of these specific details. In other instances, well-known structures and devices are shown in block diagram form to avoid obscuring the underlying principles of the present invention.

Message Filtering Apparatus

FIG. 1 depicts an exemplary message filtering apparatus 101 on which the embodiments of the invention described below may be implemented. As depicted, the message filtering apparatus includes an email analyzer 102 and a model 103. The email analyzer 102 includes an image analyzer component 104 which implements the techniques described herein to identify spam images.

When an email 105 including an image is received by the message filtering apparatus 101, as shown in FIG. 1, the email as a whole is scanned by the analyzer 102, and the image portion of the email is scanned by the image analyzer 104. The analyzer 102 determines whether the email 105 contains known spammy “features”—i.e., characteristics of an email which are known to be found in spam. For example, the analyzer 102 may determine that the email 105 includes the phrase, “get results fast!” or the word “Viagra.” Image “fingerprints” are also stored as features and identified by the image analyzer component 104.

Once the features of an email message have been identified, a mathematical model 103 is used to apply “weights” to each of the features. Features which are known to be a relatively better indicator of spam are given a relatively higher weight than other features. The feature weights are determined via “training” of classification algorithms such as Naïve Bayes, Logistic Regression, Neural Networks, etc.

The combined weights are then used to arrive at a spam “score” 108. If the score is above a specified threshold value, then the email is classified as spam and filtered out of the email stream. By contrast, if the score is below the specified value, then email is forwarded on to a user's email account.

The embodiments of the invention described below focus on the manner in which the image analyzer 104 identifies image features within email messages. It should be noted that not all of the specific operations set forth below are needed for complying with the underlying principles of the invention. Furthermore, the discussion below is not inclusive of all methods, steps, or processes covered by the present invention.

The Image Analyzer

FIG. 2 illustrates one embodiment of a method executed by the image analyzer to identify image features. The image analyzer first determines if the image is “oversized” at 201 (e.g., above a specified threshold value). Spam images are typically small so that individuals on dial-up internet connections will be able to quickly view spam in their email accounts. Moreover, given that service providers may cap the amount of information that may be sent by an individual user, spammers conserve bandwidth by using smaller images.

If the image from email 105 is oversized, then an “oversized” feature is fired by the image analyzer 104 at 202 and no further image analysis is performed. If the image is not “oversized”, then the image is pre-processed at 203. In one embodiment, pre-processing includes obtaining the image format, image width, image height and/or image size. Then the image analyzer 104 determines whether the image is in a supported format at 204 (e.g., a format which the image analyzer is capable of analyzing). Examples of supported formats are Graphic Interchange Format (“GIF”) and Joint Photographic Experts Group (“JPEG”) images. If so, the image data is read by the image processing application at 206. If the image is unsupported, then an “unsupported” feature is fired at 205 and the process terminates.

As described below, in one embodiment of the invention, the image format and image width are used as a hash key to an image fingerprint hash table.

For GIFs and JPEGs, the width is relatively straightforward to determine. Once the image is pre-processed at 203, then the image data of the image is read and analyzed at 206. In one embodiment of the present invention, ImageMagick™ is the program used to read the image data. However, various other image processing programs may be used while still complying with the underlying principles of the invention (e.g., Adobe Photoshop, Corel Draw, Paint Shop Pro).

After the image data is read at 206, the image analyzer 104 determines if the image is corrupted at 207. Spammers may “corrupt” an image by making the image unreadable by filters. For example, a spammer may change the format or embed data within the image so that the reader in 206 is unable to read the image data.

In one embodiment of the invention, if the image is corrupted, then the image analyzer 104 fires a “corrupt” feature at 208 in response to being unable to read the image data. At 209, the image analyzer 104 performs a “GIF/JPEG Feature Detection Algorithm” (hereinafter “GIF80/JPEG80”) to create a signature for the image and to search for a match for the signature. The “GIF80/JPEG80” algorithm is described in detail below.

By contrast, if the image is not corrupted, the image analyzer 104 executes a “Fuzzy Feature Detection Algorithm” (hereinafter “Fuzzy80”) to create a different signature for the image and to search for a match for the signature. In this embodiment, computing resources are conserved because the GIF80/JPEG80 algorithm is more computationally intensive than the Fuzzy80 algorithm (although the GIF80/JPEG80 algorithm is more suitable for corrupted images). Therefore, running the Fuzzy80 algorithm, if possible, over GIF80/JPEG80 algorithm saves processing power.

In one embodiment of the invention, both the Fuzzy80 algorithm and the GIF80/JPEG80 algorithm may be run in conjunction with each other in order to cross-validate the algorithm results. In yet another embodiment of the invention, only the GIF80/JPEG80 algorithm is performed on the image and the Fuzzy80 algorithm is omitted. In this embodiment, the image analyzer 104 may not need to determine if the image is corrupted 208.

Returning to the first described embodiment, shown in FIG. 2, once the image analyzer 104 performs the GIF80/JPEG80 algorithm 209 and/or the Fuzzy80 algorithm 210, at 211, the image analyzer 104 determines whether any signatures returned by the preformed algorithm match the signatures of any known spam images.

If the image analyzer 104 does not match the signatures of the current image to any signatures of the previously encountered spam images, then the image analyzer 104 fires a “No Match” feature at 212 to indicate that the image does not have any features. If, however, the image analyzer 104 matches at least one signature of the current image with at least one signature of a previously encountered spam image, then the image analyzer 104 fires the features 107 corresponding to the matched signatures 213. The features 107 may be identified by title, type, a special list, or any other technique for identifying data objects.

As previously stated, once the image features 107 are sent to the model 103, the model 103 attaches weights to the image features 107 (along with the weights of the other features of the email 105), and computes a score 108 to determine if the email 105 is spam (e.g., if the score is above a specified threshold value).

Image Signatures

A “Fuzzy80” algorithm and a “GIF80/JPEG80” algorithm were briefly mentioned above. Different embodiments of these two algorithms will now be described in detail.

An image “signature” is a unique code created by performing a hash on the image data. In one embodiment of the present invention, a Message Direct 5 (MD5) hash function is performed on the image data to create a 128-bit signature. In another embodiment of the invention, a Secure Hash Algorithm (SHA), such as SHA-1, may be used. The present invention should not be limited, though, to a specific algorithm, as almost any hash function (or one-way hash function) may be used to create the image signature.

In order to trick signature-based spam filters, spammers manipulate images in various ways including, for example:

    • randomly changing the values of unused entries in the GIF color table;
    • appending random data within the image file after the image data; and
    • randomly changing the image data in the last few rows of the image.

In contrast to prior art spam filters, the “Fuzzy80” algorithm and the “GIF80/JPEG80” algorithm described below produce a recognizable signature for images that have been manipulated by spammers.

Fuzzy Feature Detection Algorithm (Fuzzy80)

FIG. 3 illustrates one embodiment of the Fuzzy80 algorithm. At 301, the image analyzer 104 crops the image. By way of example, in FIG. 4, image 401 is cropped on all four sides of the image to yield the cropped image 402. To crop, the top image data (y-offset percentage) 403 and left image data (x-offset percentage) 404 are removed. In addition, the same amounts of data are removed from the bottom and right sides of the image, yielding an effective image of width of x-limit percentage 405 and an effective height of y-limit percentage 406. In one embodiment of the present invention, the y-offset percentage 403 and x-offset percentage 404 is twenty percent of the height and width of the image. The percentage may be any number, though, and the present invention should not be limited to any particular percentage of image cropping.

FIG. 5 illustrates another exemplary cropping method in which the left, top, and right sides of the image are not cropped. A percentage of the bottom 501 of the image 500 is cropped, leaving an effective image 500 of height 502. In one embodiment of the invention, the percentage is twenty percent of the height of the image 500. However, various other percentages may be employed while still complying with the underlying principles of the invention.

Through cropping of the bottom percentage 501 of the image 500, the width 503 of the image 500 is kept intact. The width 503 may be the original width of the image of email 105 or may be the modified width of the image after pre-processing the image in 204, illustrated in FIG. 2. In addition, the offset 504 of the image 500 is zero.

By cropping the image as described above, the effects of random data appended to the end of the image and/or modifications to the data at the end of the image are removed. Thus, performing a hash function on the remaining image data still produces a recognizable signature.

Referring back to FIG. 3, once the image is cropped, the image analyzer 104 converts the remaining image to an image of size 4×4 pixels at 302. In one embodiment of the present invention, the filtering is a down-conversion or depixelation of the image through a blurring of the image until the image is 16 pixels (4×4) in size. In one embodiment, the image is divided into quadrants, and each quadrant is again divided into quadrants. Then each smaller quadrant's image pixels would be averaged to a representative pixel value for the smaller quadrant. In another embodiment of the present invention, a Gaussian Blur may be performed on the image until 16 definable pixels exist. Various alternate techniques may be used to depixelate the image while still complying with the underlying principles of the invention.

In other embodiments, the image may be converted to a size different than a 4×4 image, such as a 10×10 image. The larger the image after conversion, the more accurate the signature results will be, but more computing resources will be necessary as more points will exist within the image. Thus, the underlying principles of the invention do not depend on the size of the image, as different applications may require different sizes.

Referring back to FIG. 3, once the image is converted, a vector is determined for the converted image at 303. For example, for a converted RGB image of size 4×4, the vector will be 48 numbers since there will be 16 red numbers, 16 blue numbers, and 16 green numbers to describe the colors of the 16 pixels of the image.

Once a vector is created for the converted image, the vector is matched against vectors of known spam images at 304. For the present embodiment of the Fuzzy80 algorithm, the vector of an image is considered the image's “signature.”

FIG. 6 illustrates one embodiment of comparing the vector of the present image against the vectors of previously encountered spam images. First, the image analyzer 104 determines if any spam images exist that are the format of the present image 601. In one embodiment, the image formats 601 are contained within a hash table indexed by format. If no spam images of the same format exist, then the image analyzer 104 does not need to compare vectors (i.e., because no match is possible if the size does not match). In the example shown in FIG. 6, the different formats are illustrated as f1 602, f2 603, f3 604, and so on.

The image analyzer 104 determines if any spam images exist that have the width of the present image 605. In one embodiment, the image widths 605 are contained within a hash table indexed by width. If no spam images of the same width exist, then the image analyzer 605 does not need to compare vectors. In the example, shown in FIG. 6, the different widths are illustrated as n1 606, n2 607, n3 608, and so on.

Then, the image analyzer 104 crops the present image for each different x-offset, y-offset, x-limit, y-limit listed in the list 609 and the vectors of spam images categorized under these x-offet, y-offset, x-limit, y-limit are compared to the vector of the cropped present image.

For example, as illustrated in FIG. 6, vector n1A 613, n1B 614 and n1C 615 with their rule ids and factors are categorized under format f1 601 and width n1 606 with a list 610 of fingerprint boundbox; vectors n2A 616, n2B 617, and n2C 618 with their rule ids and factors are categorized under format f1 602 and width n1 606 with a list 610 of fingerprint boundbox. In one embodiment of the present invention, the boundbox includes the x-offset, y-offset, x-limit, and y-limit. Thus, under the one embodiment, bb1 610 would include xoffset1, yoffset1, xlimit1, and ylimit1. bb2 611 would include xoffset2, yoffset2, xlimit2, and ylimit2. Hence, if the present image is format f1 with width n1, then the vectors n1A 613, n1B 614, n1C 615, n2A 616, n2B 617, n2C 618, and so on are compared against the present vector using the different fuzzy factors (i.e. thresholds as described below) Factor_n1A, Factor_n1B, Factor_n1C, Factor_n2A, Factor_n2B, Factor_n2C and so on.

In one embodiment, error caused by depixelation is accounted for by comparing vectors to determine if they are within a range of similarity instead of being exactly the same. For example, in one embodiment of the present invention, the image analyzer 104 determines if the absolute differences between the numbers of the two vectors are cumulatively below a specified threshold. For example, in comparing two 48 number vectors, the first number of each vector is compared, then the second number, and the third, and so on, until all numbers of both vectors are compared. Then, the absolute difference between the first numbers is added to the absolute difference between the second numbers, which is added to the absolute difference between the third numbers, etc, until all of the absolute differences are added together. If the sum is less than a specific numerical threshold, then the two vectors are considered a “match.” If the sum is greater than the threshold, then the two vectors are considered to not match. FIG. 6A illustrates a specific example of the process using two 8 number vectors. The threshold is arbitrarily set to 40 in this example.

In one embodiment of the invention, the threshold is five times the size of the vector. Therefore, on average, each position of the present vector must be within five numbers of the equivalent position of a compared vector in order for there to exist a match. In addition, in one embodiment of the invention, multiple thresholds for segments of the vectors exist, and other means of comparing the vectors may exist.

A present vector may match multiple vectors of known spam images. Therefore, the present vector is compared to all of the vectors categorized under the same list 609 When a match is determined, the features corresponding to the matching vector are determined.

As illustrated in FIG. 6, an array of features corresponding to the vectors 619 is maintained. In the illustrated example, feature Spamftr_img1_n1A 620 corresponds to vector n1A 613; feature Spamftr_img1_n1B 621 corresponds to vector n1B 614; feature Spamftr_img1_n1C 622 corresponds to vector n1C 615; and so on. In other embodiments, more than one feature corresponds to a vector. The names given to the features 620-625 illustrated in FIG. 6 may be any type of identifier to identify the feature. Thus, no special format of the feature identifier is required by the present invention.

Once all of the matching features are determined, the image analyzer sends the features as image features (213 of FIG. 2) to the model 103 in order to compute the score 108.

The underlying principles of the invention do not depend on the bit level of color. Therefore, any number of bits for color depth may be used (e.g., 16, 32, 64, 128, 256, etc., shades of red, green, and blue). In addition, any format of the pixel definition may be used, such as YMV (creating a 48 number vector for 16 pixels) or a static color table (creating a 16 bit vector for 16 pixels).

GIF/JPEG Feature Detection Algorithm (GIF80/JPEG80)

In one embodiment of the present invention, the image analyzer 104 first determines whether the present image is a GIF or a JPEG. If the image is a GIF, the image analyzer 104 runs the GIF80 algorithm (illustrated by FIG. 7). If the image is a JPEG, the image analyzer 104 runs the JPEG80 algorithm (illustrated in FIG. 10). If the image is neither GIF nor JPEG, the image analyzer may attempt to execute one of the algorithms or a separate algorithm for the different type image.

GIF80 Algorithm

FIG. 7 illustrates one embodiment of the GIF80 algorithm and FIG. 8 illustrates an exemplary format of a GIF image. The GIF format includes a header 801, which includes a six byte tag 802 to indicate if the image is a GIF, an overall width field 803, an overall height field 804, an overall color field 805, a background color field 806, and an aspect ratio field 807. The GIF further includes an overall color table 808, and extension data 809.

If the image is a “moving” image or multiple images shown in succession to simulate motion, the above information is typically universal to all of the images within the sequence of images for the “moving” image. In one embodiment, the image analyzer 104 may analyze only the first image in the stack of images. Alternatively, the image analyzer 104 may analyze any image in the stack or may analyze any number of the images in the stack.

Referring back to FIG. 8, the specific information defining the GIF image 810 includes the image identification (e.g., Image n of N) 811, the left position of the image 812, the top position of the image 813, the width of the image 814, the height of the image 815, the color field of the image 816, the color table of the image 817, and the image data 818. In one embodiment, for the image preprocessing operation 204 mentioned above, the image analyzer 104 retrieves or expands the width 814 (rather than the overall width 803).

In one embodiment of the invention, the image analyzer determines if a GIF is manipulated by comparing the overall width 803 to the width 814, the overall height 804 to the height 815, the overall color field 805 to the color field 816, and/or the overall color table 808 to the color table 817. Compared fields that do not match may indicate that the GIF has been manipulated.

Referring back to FIG. 7, the image analyzer 104 attempts to identify the GIF header 802. If the image does not have a GIF header 802, then the GIF80 algorithm is not performed on the image. If the image has a GIF header 802, then the image analyzer 104 determines if at least one signature of a spam image of the same size as the present GIF image exists. As previously described, the size may be the actual width 814 of the GIF, or the size may be the expanded width of the GIF generated through preprocessing (operation 204 of FIG. 2).

If no signature exists for a spam image of the same size as the present image, then the GIF80 algorithm is ended. If, however, a signature exists for a spam image of the same size as the present image, then the image analyzer 104 crops the image 703. The image may be cropped as previously discussed under the Fuzzy80 algorithm and illustrated in FIGS. 4 and 5. If the image is cropped as per FIG. 5, a portion 820 of the image data 818 will be cropped out, leaving the effective image data 818 shown in FIG. 8.

Once the image is cropped, the image analyzer determines the signature for the image data 704. As previously described, the signature is determined by performing a hash (e.g., MD5 algorithm) on the effective image data 819 of the GIF. Once the signature is determined for the image, the image analyzer 104 compares the determined signature to signatures of previously encountered spam images to identify features of the present image.

FIG. 9 illustrates one embodiment of the image analyzer 104 comparing the determined signature of the present image to stored signatures of previously encountered spam images. In the embodiment illustrated in FIG. 9, a signature is derived by performing the MD5 hash algorithm on an image.

First, the image analyzer 104 compares the width of the present image to the widths used to categorize the stored signatures 901. As described above, the signatures may be stored in the form of a hash table indexed by width. The widths illustrated are n1 902, n2 903, and n3 904, etc. The image analyzer 104 uses the image width array 901 to determine if the present image width is the same width as any previously encountered spam images. Alternative to width, other size measurements may be used to categorize stored signatures 901, such as height and overall number of pixels.

As illustrated in FIG. 9, groups of MD5 signatures 905 are identified via entries in the image width hash table 901 and lengths of N bytes 937 of the corresponding widths over which the MD5 signatures 905 are computed. For example, MD5 signature n1A 906, MD5 signature n1B 907, and MD5 signature n1C 908 correspond to byte length N of n1A 928, n1B 929, and n1C 930, respectively, of image width n1 902; MD5 signature n2A 909, MD5 signature n2B 910, and MD5 signature n2C 911 correspond to byte length N of n2A 931, n2B 932, and n2C 933, respectively, of image width n2 903; and MD5 signature n3A 912, MD5 signature n3B 913, and MD5 signature n3C 9013 correspond to byte length N of n3A 934, n3B 935, and n3C 936, respectively, of image width n1 904. As in the Fuzzy80 algorithm, any number of signatures may correspond to each width in the Image Width array 901.

For the GIF80 algorithm, the determined signature of the present image may be compared with spam signatures for an exact match.

By way of example, and not limitation, if the present image is of size n1 902, the signature of the present image may be compared to all of the signatures 906-908, etc, identified by that size. In another embodiment of the present invention, the image analyzer 104 may stop comparing signatures once an exact match is found.

Once all of the matches for the signature of the present image have been determined, the features 915 corresponding to the matching signatures are fired by the image analyzer 104 and corresponding weights are assigned by the model 103. In the specific example of FIG. 9, features Spam_ftr_img1_n1A 916 and Spam_ftr_img2_n1A 917 correspond to MD5 signature n1A 906, and so on. FIG. 9 also illustrates that multiple features may be associated with one signature.

JPEG80

Referring to FIG. 10, the JPEG80 algorithm is similar to the GIF80 algorithm, the primary different being that the GIF80 algorithm categorizes signatures by image width while the JPEG80 algorithm categorizes signatures by image compression type. FIG. 11 illustrates a lookup table for JPEG images which is similar to the lookup for GIFs (FIG. 9) but which categorizes JPEGs based on compression type.

In other embodiments of the present invention, the image analyzer 104 categorizes signatures of JPEGs by image width.

FIG. 12 illustrates a series of data structures employed by both the GIF80 algorithm and the JPEG80 algorithm. A first lookup table 1201 categorizes images by image type. The entries in the first lookup table point to other sub-tables which are indexed based on different image features (e.g., width, compression type, etc). For example, if the present image is a GIF, then a lookup sub table 1204 indexed by image width is identified (e.g., similar to the lookup table in FIG. 9). Similarly, if the present image is a JPEG, then the lookup sub table (not shown) may be the lookup table of FIG. 11.

Embodiments of the invention may include various steps as set forth above. The steps may be embodied in machine-executable instructions which cause a general-purpose or special-purpose computer processor to perform certain steps. Alternatively, these steps may be performed by specific hardware components that contain hardwired logic for performing the steps, or by any combination of programmed computer components and custom hardware components.

Elements of the present invention may also be provided as a machine-readable medium for storing the machine-executable instructions. The machine-readable medium may include, but is not limited to, floppy diskettes, optical disks, CD-ROMs, and magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, magnetic or optical cards, propagation media or other type of media/machine-readable medium suitable for storing electronic instructions. For example, the present invention may be downloaded as a computer program which may be transferred from a remote computer (e.g., a server) to a requesting computer (e.g., a client) by way of data signals embodied in a carrier wave or other propagation medium via a communication link (e.g., a modem or network connection).

Throughout the foregoing description, for the purposes of explanation, numerous specific details were set forth in order to provide a thorough understanding of the invention. It will be apparent, however, to one skilled in the art that the invention may be practiced without some of these specific details. Moreover, the underlying principles of the invention may be implemented within virtually any type of computing platform including standalone personal computer configurations and server configurations.

Claims

1. A computer-implemented method for determining spam features of a present image, comprising:

determining whether a format of the present image is in a supported format of an image analyzer;
reading the present image by an image reader of the image analyzer when the format of the present image is determined to be in the supported format;
analyzing the present image by the image analyzer to determine whether the present image is corrupted, wherein the present image is determined to be corrupted when the image reader of the image analyzer determines that the format of the present image has been changed;
selecting between at least a first feature detection algorithm and a second feature detection algorithm based on the determination of whether the present image is corrupted;
when the present image is determined not to be corrupted, executing the first feature detection algorithm by the image analyzer to create a first signature to determine if the present image is a spam image, wherein the first feature detection algorithm comprises depixelating at least a first portion of the present image and processing the portion of the present image after the depixelation to create the first signature, wherein processing the portion of the present image comprises comparing a vector associated to the portion of the present image with a set of vectors of known spam images, and determining if the vector is within a range of similarity with one of the set of vectors; and
when the present image is determined to be corrupted, executing the second feature detection algorithm by the image analyzer to create a second signature to determine if the present image is a spam image, wherein the second feature detection algorithm comprises analyzing at least a second portion of the present image to create the second signature, and comparing the second signature to stored signatures of known spam images.

2. The computer-implemented method as in claim 1 wherein, if the first feature detection algorithm is executed and does not detect that the image is a spam image, then executing the second feature detection algorithm to validate the determination of the first feature detection algorithm.

3. The computer implemented method as in claim 2 wherein the second feature detection algorithm computes a precise signature of the second portion of the present image and wherein the first feature detection algorithm computes a signature of the first portion of the present image which is relatively less precise than the signature calculated by the second feature detection algorithm.

4. The computer-implemented method as in claim 1 wherein the first feature detection algorithm comprises creating the first signature for the present image after the depixelation and searching for a match for the first signature, and wherein the second feature detection algorithm comprises creating the second signature for the present image and searching for a match for the second signature.

5. The computer-implemented method as in claim 1 wherein the first portion of the present image is selected through cropping, wherein the cropping removes a percentage of at least one side of the present image.

6. The computer-implemented method as in claim 1 wherein the depixelation comprises blurring of the first portion of the present image until the present image reaches a certain size.

7. The computer-implemented method as in claim 4 wherein the first signature for the present image is the vector for the first portion of the present image.

8. A non-transitory machine-readable medium having machine-executable instructions stored thereon, which when executed by a computer processing system, causes the computer processing system to perform a method for determining spam features of a present image, the method comprising:

determining whether a format of the present image is in a supported format of an image analyzer;
reading the present image by an image reader of the image analyzer when the format of the present image is determined to be in the supported format;
analyzing the present image by the image analyzer to determine whether the present image is corrupted, wherein the present image is determined to be corrupted when the image reader of the image analyzer determines that the format of the present image has been changed;
selecting between at least a first feature detection algorithm and a second feature detection algorithm based on the determination of whether the present image is corrupted;
when the present image is determined not to be corrupted, executing the first feature detection algorithm by the image analyzer to create a first signature to determine if the present image is a spam image, wherein the first feature detection algorithm comprises depixelating at least a first portion of the present image and processing the portion of the present image after the depixelation to create the first signature, wherein processing the portion of the present image comprises comparing a vector associated to the portion of the present image with a set of vectors of known spam images, and determining if the vector is within a range of similarity with one of the set of vectors; and
when the present image is determined to be corrupted, executing the second feature detection algorithm by the image analyzer to create a second signature to determine if the present image is a spam image, wherein the second feature detection algorithm comprises analyzing at least a second portion of the present image to create the second signature, and comparing the second signature to stored signatures of known spam images.
Referenced Cited
U.S. Patent Documents
5465353 November 7, 1995 Hull et al.
5721788 February 24, 1998 Powell et al.
5872865 February 16, 1999 Normile et al.
5893095 April 6, 1999 Jain et al.
5898779 April 27, 1999 Squilla et al.
5899999 May 4, 1999 De Bonet
5913205 June 15, 1999 Jain et al.
5937084 August 10, 1999 Crabtree et al.
5963670 October 5, 1999 Lipson et al.
6175829 January 16, 2001 Li et al.
6307955 October 23, 2001 Zank et al.
6389424 May 14, 2002 Kim et al.
6415282 July 2, 2002 Mukherjea et al.
6779021 August 17, 2004 Bates et al.
6847733 January 25, 2005 Savakis et al.
6931591 August 16, 2005 Brown
7089241 August 8, 2006 Alspector et al.
7437408 October 14, 2008 Schwartz et al.
7475061 January 6, 2009 Bargeron et al.
7624274 November 24, 2009 Alspector et al.
7715059 May 11, 2010 Advocate et al.
7716297 May 11, 2010 Wittel et al.
7817861 October 19, 2010 Lee
8103048 January 24, 2012 Sheinin et al.
8175387 May 8, 2012 Hsieh et al.
8290203 October 16, 2012 Myers et al.
8290311 October 16, 2012 Myers et al.
8356076 January 15, 2013 Wittel et al.
8489689 July 16, 2013 Sharma et al.
8792728 July 29, 2014 Tang
20010004739 June 21, 2001 Sekiguchi et al.
20010039563 November 8, 2001 Tian
20020032672 March 14, 2002 Keith, Jr.
20020076089 June 20, 2002 Muramatsu
20020180764 December 5, 2002 Gilbert et al.
20030053718 March 20, 2003 Yamamoto
20030123737 July 3, 2003 Mojsilovic et al.
20030174893 September 18, 2003 Sun et al.
20030215135 November 20, 2003 Caron et al.
20040148330 July 29, 2004 Alspector et al.
20040153517 August 5, 2004 Gang et al.
20040177110 September 9, 2004 Rounthwaite et al.
20040213437 October 28, 2004 Howard et al.
20040221062 November 4, 2004 Starbuck et al.
20040260776 December 23, 2004 Starbuck et al.
20050030589 February 10, 2005 El-Gazzar et al.
20050050150 March 3, 2005 Dinkin
20050060643 March 17, 2005 Glass et al.
20050076220 April 7, 2005 Zhang et al.
20050097179 May 5, 2005 Orme
20050120042 June 2, 2005 Shuster et al.
20050120201 June 2, 2005 Benaloh
20050216564 September 29, 2005 Myers et al.
20060092292 May 4, 2006 Matsuoka et al.
20060093221 May 4, 2006 Kasutani
20060123083 June 8, 2006 Goutte et al.
20060168006 July 27, 2006 Shannon et al.
20060168041 July 27, 2006 Mishra et al.
20060184574 August 17, 2006 Wu et al.
20060190481 August 24, 2006 Alspector et al.
20070130350 June 7, 2007 Alperovitch et al.
20070130351 June 7, 2007 Alperovitch
20070211964 September 13, 2007 Agam et al.
20080002914 January 3, 2008 Vincent et al.
20080002916 January 3, 2008 Vincent et al.
20080063279 March 13, 2008 Vincent et al.
20080091765 April 17, 2008 Gammage et al.
20080127340 May 29, 2008 Lee
20080130998 June 5, 2008 Maidment et al.
20080159585 July 3, 2008 True et al.
20080159632 July 3, 2008 Oliver et al.
20080175266 July 24, 2008 Alperovitch et al.
20080177691 July 24, 2008 Alperovitch et al.
20080178288 July 24, 2008 Alperovitch et al.
20080219495 September 11, 2008 Hulten et al.
20090043853 February 12, 2009 Wei et al.
20090077476 March 19, 2009 Hua et al.
20090100523 April 16, 2009 Harris
20090110233 April 30, 2009 Lu et al.
20090113003 April 30, 2009 Lu et al.
20090141985 June 4, 2009 Sheinin et al.
20090220166 September 3, 2009 Choi et al.
20090245635 October 1, 2009 Yehezkel et al.
20100158395 June 24, 2010 Sathish et al.
20110052074 March 3, 2011 Hayaishi
20110244919 October 6, 2011 Aller et al.
20120087586 April 12, 2012 Sheinin et al.
20120240228 September 20, 2012 Alperovitch et al.
20130212090 August 15, 2013 Sperling et al.
Foreign Patent Documents
2005011325 January 2005 JP
Other references
  • Yan Gao, Ming Yang and Xiaonan Zhao, “Image Spam Hunter”, Jun. 15, 2006, http://www.cs.northwestern.edu/˜yga751/ML/ISH.htm, p. 1-8.
  • Author Unknown, “Tumbleweed advances the war on spam with new adaptive image filtering technology”, Nov. 7, 2006, http://www.axway.com/press-releases/2006/tumbleweed-advances-war-spam-new-adaptive-image-filtering-technology, p. 1-2.
  • Garrettson, Cara, “IronPort: Image-spam filters produce high catch rate”, Nov. 17, 2006, “http://www.computerworld.com.au/article/166513/ironport_image-spam_filters_produce_high_catch_rates/”, p. 1-3.
  • Corman, T, et al., “Introduction to Algorithms”, 2000, McGraw-Hill Book Company, p. 221-226.
  • “The New Wave of Spam”, http://www.symantec.com/business/resources/articles/article.jsp?aid=new_wave_of_spam, (May 7, 2007).
  • Aradhye, H.B. , et al., “Image Analysis for Efficient Categorization of Image-Based Spam e-mail”, Proceedings of the 2005 Eighth International Conference on Document Analysis and Recognition (ICDAR '05), (2005).
  • Burton, Brian, “Spam Probe—Bayesian Spam Filtering Tweaks”, http://web.archive.org/web/20060207003157/http://spamprobe.sourceforge.net/, (2006), 1-5.
  • Johnson, Neil F., “In Search of the Right Image: Recognition and Tracking of Images in Image Databases, Collections, and the Internet”, Center for Secure Information Systems Technical Report CSIS-TR-99-05NFJ., (Apr. 1999), 1-13.
  • Schwartz, Randal , “Finding Similar Images, Perl of Wisdom”, Linux Magazine, Aug. 14, 2003, www.linux-mag.com/index2.php?option=com_content&task=view&id=Itemid=221, (Aug. 15, 2003), 6 pages.
  • Wang, Zhe, et al., “Filtering Image Spam with Near-Duplicate Detection”, CEAS 2007—The Third Conference on Email and Anti-Spam, 2007, (2007), 1-10.
  • Wu, C.T., et al., “Using Visual Features for Anti-Spam Filtering”, IEEE International Conference on Image Processing, vol. 3, (2005), 4 pages.
  • Notice of Allowance, U.S. Appl. No. 11/652,715, dated Jun. 12, 2012, 9 pages.
  • Final Office Action, U.S. Appl. No. 11/652,715, dated Nov. 28, 2011, 25 pages.
  • Non-Final Office Action, U.S. Appl. No. 11/652,715, dated Mar. 10, 2011, 35 pages.
  • Final Office Action, U.S. Appl. No. 11/652,715, dated Nov. 23, 2010, 25 pages.
  • Non-Final Office Action, U.S. Appl. No. 11/652,715, dated Jun. 11, 2010, 31 pages.
  • Notice of Allowance, U.S. Appl. No. 11/652,716, dated Jun. 13, 2012, 9 pages.
  • Final Office Action, U.S. Appl. No. 11/652,716, dated Nov. 28, 2011, 9 pages.
  • Non-Final Office Action, U.S. Appl. No. 11/652,716, dated Mar. 18, 2011, 34 pages.
  • Final Office Action, U.S. Appl. No. 11/652,716, dated Oct. 28, 2010, 41 pages.
  • Non-Final Office Action, U.S. Appl. No. 11/652,716, dated May 12, 2010, 30 pages.
Patent History
Patent number: 10095922
Type: Grant
Filed: Oct 11, 2012
Date of Patent: Oct 9, 2018
Patent Publication Number: 20130039582
Assignee: Proofpoint, Inc. (Sunnyvale, CA)
Inventors: John Gardiner Myers (Santa Clara, CA), Yanyan Yang (Sunnyvale, CA)
Primary Examiner: Gandhi Thirugnanam
Application Number: 13/650,107
Classifications
Current U.S. Class: Classification (382/224)
International Classification: G06K 9/54 (20060101); G06K 9/00 (20060101); G06Q 10/10 (20120101);